fix(agent-core-v2): pre-shrink compaction history to the effective model window

[Grapedge]

Related Issue

No tracking issue — internal spec-driven fix (compaction pre-shrink).

Problem

After accumulating a conversation larger than a small-window model can hold (e.g. ~300k tokens on a 1M-window model), switching to a smaller-window model (e.g. 256k) makes compaction fail: the first summary request sends the entire oversized history to the current model, which is guaranteed to overflow. Whether compaction recovers depended entirely on the provider's error text being classified as context overflow — otherwise the first request failed directly with compaction.failed, leaving the session stuck (context still oversized, every next message re-triggering the same failing auto-compaction chain).

This affected all compaction triggers (manual /compact, pre-step water level, post-step check, turn overflow recovery) since they share one summary path.

What changed

Before sending the first summary request, compaction now proactively computes an input budget from the current model's effective context window and pre-shrinks the history to fit:

budget = floor((effective window − output reserve) × 0.85) − request overhead

• Effective window: existing getEffectiveMaxContextTokens() (min of configured capability and the per-alias observed overflow learning value); unknown window (≤ 0) skips pre-shrink entirely, preserving current behavior.
• Output reserve: min(compaction output cap, floor(window / 8)) so the summary output is never clamped to zero; request overhead: the existing requestTokens([]) estimate (system prompt + tools); safety factor: the existing OVERFLOW_CONTEXT_SAFETY_RATIO.
• Trimming reuses the existing "keep recent messages within token budget" function that backs the reactive overflow shrink (one trimming implementation, two call sites); dropped messages are accounted into the existing droppedCount field.
• The reactive 0.7/0.5/0.35 shrink chain stays as the fallback when estimation error still overflows. Summary text, telemetry schema, result shape, TUI and app clients are unchanged.

One existing test (auto-compacts very large context in one full-history round) asserted the old "first request always sends full history" behavior for an over-window context; it now asserts the intended pre-shrunk first request.

Checklist

• I have read the CONTRIBUTING document.
• I have linked a related issue (external PRs: the issue must have a maintainer's /approve).
• I have added tests that prove my feature works.
• Ran gen-changesets skill, or this PR needs no changeset.
• Ran gen-docs skill, or this PR needs no doc update.

[changeset-bot]

🦋 Changeset detected

Latest commit: b6ad5d4

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 1 package

Name	Type
@moonshot-ai/kimi-code	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

[pkg-pr-new]

pnpm dlx https://pkg.pr.new/@moonshot-ai/kimi-code@b6ad5d4

npx https://pkg.pr.new/@moonshot-ai/kimi-code@b6ad5d4

commit: b6ad5d4

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: b6ad5d4b3c

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[Copilot]

🔵 Needs a closer look

It changes core compaction request-sizing behavior in the agent engine affecting reliability and cost for all long sessions, relying on heuristic token-budget math that warrants final human review.

Pull request overview

This PR fixes a compaction failure that occurs after switching from a large-context model to a smaller-context one: the first summary request previously always sent the entire (now over-window) history, which was guaranteed to overflow and could leave the session stuck. The change proactively pre-shrinks the history to an input budget derived from the current model's effective window before issuing the first summary request, while leaving the existing reactive overflow-shrink chain as a fallback. This is a core agent-core-v2 compaction-path change that surfaces to CLI users as improved reliability.

Changes:

• Add preShrinkHistoryToWindowBudget and call it before the compaction request loop, folding any pre-dropped messages into the existing droppedCount.
• Budget = floor((effectiveWindow − outputReserve) × 0.85) − requestOverhead, reusing the existing "keep recent messages within token budget" trimmer; unknown window (≤ 0) skips pre-shrink.
• Update one existing test to assert the pre-shrunk first request and add a new FullCompaction pre-shrink suite; add a patch changeset.

File summaries

File	Description
packages/agent-core-v2/src/agent/fullCompaction/fullCompactionService.ts	Adds the pre-shrink budget computation and applies it to the summary history before the first request, updating droppedCount.
packages/agent-core-v2/test/agent/fullCompaction/fullCompaction.test.ts	Updates the over-window test and adds a suite covering pre-shrink, budget/overhead accounting, dropped-count, reactive fallback, unknown window, and oversized-tail cases.
.changeset/compaction-pre-shrink.md	Adds a patch changeset for @moonshot-ai/kimi-code with a single user-facing sentence.

Review details

• Files reviewed: 3/3 changed files
• Comments generated: 0
• Review effort level: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.